AI agent contracts: repricing as token costs fall

AI agent contracts are the wrong place to lock a rate for three years. On 9 November 2026, Anthropic cut Fable from $10 to $8 per million input tokens with the Fable 5.2 release. Meta launched Muse Ember at 10 cents per million input tokens for the repetitive work agents do all day (TechCrunch, TechCrunch). The raw material under every agent workflow is getting cheaper by the week.

That does not mean your bill will fall. Model prices and vendor prices move on different clocks. The gap between them is where procurement budgets leak. The fix is contractual, not technical. Build price-review clauses, benchmark rights and unit definitions into AI agent contracts before you sign.

Why cheaper tokens have not shown up on your invoice

Model makers charge per million tokens. Your agent vendor charges per user, per run, per credit or per resolution. It pays the model maker out of that price. If the fee is flat, its input costs can halve while your fee stays exactly where it is.

Several cuts landed in 2026. OpenAI cut GPT-5.6 Luna output pricing by 80% on 30 July, to $1.20 per million tokens, per Kent Applied. Anthropic's Claude Opus 5.5 arrived at $4 input and $20 output per million tokens, against $5 and $25 for Opus 5, on 22 September. Lower input prices alone do not reach you unless AI agent contracts pass them through.

Volume is the other half of the story. Falling unit prices can sit alongside a rising bill. Vercel's gateway data showed estimated spend about 13 times higher in September 2026 than in December 2025. Average price per token fell 47% over the same period (Kent Applied). Your governance and renegotiation clause has to cover both directions.

The unit on the invoice is the real price term

A pricing census of 41 agent products found 23, or 56%, billing per credit, per run or per minute as of 4 October 2026. The median entry plan was $20 a month (Gravity AI). Only four still charged per seat. Usage meters have won, and a meter is only as fair as its definition.

Vendors also change the unit rather than the number. Of 38 dated pricing changes logged since August 2024, just five were plain increases. Another 26 restructured what a plan includes or added a new unit (Gravity AI). A renewal that looks flat can cost more because a run now means less work.

Procurement should pin down:

• What counts as a run, credit, step or resolution, in plain words.
• Whether retries, failed calls and verification loops are billed.
• How cached input is priced, since it is cheaper but still charged.
• What happens when the vendor switches models mid-term.

Tie the definition to the invoice, not to a marketing page. Treat any vendor definition of a billable outcome as a contract term, because that is what it is.

Build a price-review clause that actually moves

A price-review clause should state when the price is reopened, on what evidence, and what happens if the two sides disagree. Vague language about "periodic review" gets you a meeting, not a reduction.

Work through these steps:

1. 1Set a review trigger. Calendar-based (every two quarters) is simple. Cost-based is sharper: tie it to published list prices falling by a set percentage on named models.
2. 2Name the reference models. If your agent runs on a small model for execution and a larger one for planning, list both. Ember was built to pair with a heavier model in an agentic harness. The heavier model plans and the light one handles volume (TechCrunch).
3. 3Define the evidence. Accept public price cards, vendor changelogs and benchmark reports as proof.
4. 4Agree the sharing mechanism. Either a pass-through of the cost delta or a fixed percentage reduction, with a floor and a cap.
5. 5Set the fallback. Use a third-party benchmark, or a right to terminate without penalty if no agreement is reached in 30 days.

This is the same discipline you apply to any fixed price, time and materials or dedicated team engagement. The model changes; the principle holds. Price risk should sit with the party that can manage it. Good AI agent contracts name that party in writing.

Benchmark rights: what to ask for and how to use them

Benchmark rights let you test the price after signing. Without them, you are negotiating from memory while the vendor negotiates from its rate card.

Ask for the right to run a benchmark once per contract year at your cost. Ask for results to be shared, with a renegotiation window. Define the benchmark as cost per correctly completed task, not cost per token. A task-level measure survives model swaps and catches workflows that quietly consume more calls.

Be realistic about the workload. McKinsey's July 2026 report covered over 20 production agentic coding workflows. It put roughly 60% of an agentic task's cost in response refinement rather than the first inference call, per Forkast. That is consultancy analysis on a small coding sample, not a universal rule. It still explains why cheap tokens can produce expensive agents.

Use the benchmark to ask three questions. Are we on the cheapest model that meets quality? Are verification loops controlled? Has the rate come down in line with the market? Build the answers into your AI agents evaluation so cost and quality are judged together. Benchmark rights are the part of AI agent contracts that most buyers forget to ask for.

Where savings actually leak

The biggest leak is often the model tier. A workflow that once needed a frontier model may now run on a small one. AI agent contracts should allow the vendor to switch without restarting procurement.

Anthropic says its tests show a 40% lower cost than the prior model on typical workloads. The reason is a lower price per token plus fewer tokens per task. That is the vendor's own estimate for its own model (Kent Applied).

The second leak is temporary pricing. Some cuts are promotional. OpenAI's page says GPT-5.6 Sol's promotional price is available at least through 21 November 2026. It also charges 10% more for regional processing endpoints on models released on or after 5 March 2026 (Kent Applied). If your data rules require a region, check whether that surcharge is inside your rate.

The third leak is the refinement loop. If the vendor bills per step, every retry is revenue. Cap steps per task, add an exception report, and put the cap in writing. Review it alongside your AI agent permissions controls so cost and access governance move together. AI agent contracts that ignore the loop leave the biggest savings on the table.

How to renegotiate at renewal

Come to renewal with data, not opinion. Pull three things: cost per completed task for the last two quarters, the share of spend on each model tier, and the vendor's current rate card against the one you signed.

Then run a short, structured negotiation. Offer term length in exchange for a price-review clause and benchmark rights, not in exchange for a fixed discount. Ask for a model-tier commitment and a step cap. Ask what happens if a vendor replaces a model mid-term with one that is significantly cheaper.

Where you cannot get movement, widen the field. Run a competitive round against other vendors rather than accepting the renewal rate. Vendors responding to open work through a listing of current RFPs can price against today's cost base. That gives you a real comparison point.

Frequently asked questions

Are AI agents getting cheaper in 2026?

Yes, at the model layer. Meta launched Muse Ember at 10 cents per million input tokens, and Anthropic cut Fable input pricing from $10 to $8 per million tokens, both in November 2026. Whether your bill falls depends on whether your AI agent contracts pass those cuts through.

What is a price-review clause in an AI agent contract?

It is a clause that says when the price is reopened, what evidence counts as proof of a market move, and what happens if the parties disagree. It protects the buyer when underlying model costs fall, and it protects the vendor when they rise.

Should I buy AI agents on a fixed price or a usage meter?

Usage meters are now the norm. 56% of 41 agent products in one census billed per credit, run or minute as of October 2026. Usage pricing can be fair, but only if the unit is clearly defined and capped, with review and benchmark rights in place.

What is the best metric to benchmark AI agent pricing?

Cost per correctly completed business task. Per-token prices move too fast and too independently of vendor prices. A task-level figure tracks what you actually pay for work delivered, so make it the benchmark in your AI agent contracts.

What this means for buyers

Prices at the model layer are falling faster than most contracts can follow. Treat that gap as a negotiable item, not bad luck.

Next steps:

• Audit every AI agent contract for an existing price-review trigger and benchmark right. Most will not have one.
• Rewrite the unit definitions so retries, verification loops and cached input are explicit.
• Add a model-tier commitment clause, so savings from cheaper execution models reach you.
• Run a benchmark at cost per completed task and bring the results to renewal.
• Compare at least two competing proposals before renewing, so you know the market rate rather than the incumbent's rate.

Falling token costs are good news only for buyers who wrote AI agent contracts to capture them.